故事 · 啤酒厂里诞生的“学生”分布 Origin Story · The "Student" Distribution Born in a Brewery
1908 年，都柏林 Guinness 啤酒厂的化学家 William Gosset 要用很少的几次实验判断原料质量， 却发现样本小时正态近似不准。他推导出一种尾巴更厚的分布来补偿“用 s 估 σ”的不确定。 因公司禁止员工署名发表，他用笔名"Student"刊出 —— 这就是 Student's t 分布的由来。 样本越小，t 的尾巴越厚；当 n 很大，t 就收敛回正态。小样本均值检验，从此有了正确的工具。 In 1908, William Gosset, a chemist at the Guinness Brewery in Dublin, had to judge raw-material quality from only a handful of trials — and discovered that the normal approximation breaks down on small samples. He worked out a heavier-tailed distribution to compensate for the uncertainty of "using s to estimate σ". Since the brewery forbade employees from publishing under their own names, he signed the paper "Student" — and that's where Student's t-distribution got its name. The smaller the sample, the heavier t's tails; as n grows large, t converges back to the normal. Small-sample mean testing finally had its proper tool.

1 t 落在分布哪里？冲进拒绝域了吗 Where Does t Land — Inside the Rejection Region?

df = 11

2 为什么小样本用 t 而非 z：尾巴更厚 Why Small Samples Use t, Not z: Heavier Tails

蓝线是标准正态 z，红线是当前 df 的 t 分布。df 越小，t 的尾巴越厚、中间越矮 —— 这正是对“用 s 估 σ”的不确定性补偿。n 大到几十以后，两条几乎重合。 Blue is the standard normal z; red is the t distribution at the current df. The smaller the df, the heavier t's tails and the shorter its peak — exactly the compensation for using s in place of σ. By the time n reaches a few dozen, the two curves are practically on top of each other.

3 现实里的单样本 t 检验 One-Sample t in the Real World

均值检验：一批零件的平均尺寸是否真偏离了图纸标称值，用单样本 t 一测便知。 Mean check: is the average dimension of a batch of parts really off the nominal value on the drawing? A one-sample t-test settles it.
与目标对比：实测平均交期 vs 承诺交期、平均含量 vs 规格中值，都是它的主场。 Compare to a target: measured lead time vs promised lead time, mean assay vs spec midpoint — classic one-sample t territory.
t vs z：σ 已知或大样本用 z，σ 未知且小样本用 t —— 选错会高估显著性。 t vs z: use z when σ is known or n is large; use t when σ is unknown and n is small — pick the wrong one and you overstate significance.
p 值判定：p < α 才拒绝 H₀，但别只看 p，配合效应量与置信区间才完整。 Reading the p-value: only reject H₀ when p < α — but don't stop at p; pair it with effect size and a confidence interval for the full story.
一句话In One Line
t 统计量是个信噪比：分子 x̄ − μ₀ 是“信号”（差异有多大）， 分母 s/√n 是“噪声”（这点差异在抽样里有多正常）。t 大，说明信号压过了噪声，差异不像是偶然。 因为 σ 未知、只能用 s 估，这份额外的不确定让参照分布从正态换成尾巴更厚的 t， 自由度 df = n−1 越小尾巴越厚、越保守。p 值就是“纯波动能造出这么大 t 的概率”，小到一定程度，我们才说差异是真的。 The t-statistic is a signal-to-noise ratio: the numerator x̄ − μ₀ is the "signal" (how big the gap is), the denominator s/√n is the "noise" (how normal a gap that size looks under random sampling). A large t means signal has overwhelmed noise — the gap doesn't look accidental. Because σ is unknown and we estimate it with s, that extra uncertainty swaps the reference curve from the normal to the heavier-tailed t; the smaller the df = n − 1, the heavier the tails and the more conservative the test. The p-value is "the chance pure jitter could fabricate a t this large" — small enough, and we call the gap real.
常见误用Common Mistakes
小样本、σ 未知还用 z 检验。该用 t，否则尾巴太薄、p 偏小，会高估显著性。 Running a z-test on a small sample when σ is unknown. Use t — z has tails too thin, so the p-value comes out artificially small and you overstate significance.
把 p 值当成“差异大小”。p 只说“是否偶然”，差异大小看 x̄−μ₀ 和效应量。 Reading the p-value as a measure of "how big" the difference is. p only tells you "is it chance?". For magnitude, look at x̄ − μ₀ and an effect size.
不看前提就硬套。单样本 t 假设数据近似正态、独立；强偏态/有离群点要先处理。 Applying the formula without checking assumptions. One-sample t assumes approximately normal, independent data — handle strong skew or outliers first.

单样本 t 检验